Reliable trajectory prediction in scene fusion based on spatio-temporal Structure Causal Model

Existing methods for trajectory prediction predominantly employ scene fusion to enhance model performance. However, they fail to provide a rational explanation as to why the fusion of the scene context and trajectories improves model performance, which prevents them from identifying the fundamental factors limiting model performance. Hence, this paper introduces a Structured Causal Model for trajectory prediction based on causal inference, which elucidates the genuine reasons for the performance enhancement brought about by the scene context in trajectory prediction and analyzes the confounding path interference that curtails model performance. Specifically, this paper first employs the front-door criterion to eliminate the confounders during the feature extraction process, allowing the model to fairly incorporate the scene context into the spatio-temporal state. Subsequently, a spatio-temporal causal graph is generated to further extract the causal relationship of the trajectory in the current scene, serving as the spatio-temporal representation. Finally, the technique of counterfactual representation inference extrapolates the spatio-temporal features of the historical trajectory into future traffic scenes for trajectory prediction. The effectiveness and reliability of this proposed end-to-end method has been experimentally validated on two real-world datasets in real traffic scenarios, particularly in scenarios involving interactions between multiple agents.